Meta-learned ID Embeddings for Online Inductive Recommendation

Jingyu Peng,Le Wu,Peijie Sun,Meng Wang

DOI: https://doi.org/10.1007/978-3-030-88189-4_3

2021-01-01

Abstract:Learning accurate user and item ID embeddings from user-item historical records has shown great success for recommender systems. Most of these embedding learning models are transductive and work well for users that appear in the training stage. However, in the model serving stage, new users continue to join the system. It’s important to quickly adapt to new users’ preferences for online inductive recommendation scenarios. Some previous works adopted embedding retraining or fed content data for new user ID embedding learning, these models either suffered from slow convergence or relied on auxiliary data. In this paper, we propose a meta-learned ID embedding framework for new users without using any side information in online inductive recommendation scenarios. Our key idea is that, we treat each user’s ID embedding learning as a separate task, and propose to meta-learn the initial embedding by modeling the global knowledge from all users (tasks). Each user’s embedding is initialized by the learned global knowledge instead of randomly initialization. Therefore, we could quickly adapt to a new user’s ID embedding based on a few updates from her online records, which can facilitate fast online recommendation. Moreover, our main technical contribution lies in how to learn the global prior knowledge for informative ID embedding initialization without any side information. Finally, extensive experimental results on three real-world datasets clearly show both the efficiency and effectiveness of the meta-learned ID embeddings for inductive recommendation.